ORA-600 / JAS-MIN / 0x600
JAS-MIN + ORA-600. Oracle performance audits: data, evidence, internals and an action plan. No fortune-telling from a single chart.
Your database needs people who understand what they are looking at. JAS-MIN — JSON AWR & Statspack Miner — is an Oracle performance analysis tool we develop. Our audits combine its capabilities with DBA and developer experience and knowledge of the database engine.
01 / DISTIL THE NOISE
We do not feed a pile of AWRs into a magic box. JAS-MIN organises measurements, models relationships and helps choose where to look deeper. This is how we distil signal from noise.
Measurement windows, load, waits and gaps. An empty entry does not always mean zero.
Ask how changes in statistics relate to load changes between adjacent windows. Not just which counter was largest.
Centring and standardisation make differently scaled inputs comparable. Coefficients are then converted back to original units.
Four parallel perspectives. Not four successive filters.
Ridge / MODEL NOTES
Two statistics move almost together? Huge opposing coefficients can fit that history. Ridge penalises coefficient size to reduce sensitivity to those patterns.
This stabilises the fit. It does not prove which statistic causes the problem.
Elastic Net / MODEL NOTES
Combines coefficient shrinkage with the ability to set some coefficients to zero. It helps simplify a model when there are many inputs and some tell similar stories.
A zero coefficient is not a certificate of innocence. It depends on the data, settings and other predictors.
Huber / MODEL NOTES
Small errors receive the usual squared-error treatment. Large errors exert less pressure on the fit, so a single extreme observation need not dominate the entire picture.
It does not erase incidents. We still investigate extremes separately, with their context.
Quantile 95 / MODEL NOTES
Rather than targeting the centre of the response, it models its conditional 95th percentile: a high level of load change for the given inputs. Underestimation is penalised more strongly than overestimation.
Q95 ≠ Ridge × P99. Convergence comes first. In the course sample Q95 did not converge and was excluded from qualification.
We check convergence, data quality and fit limitations. An ineligible result does not get a vote just because it looks convincing.
// FROM LEAD TO DIAGNOSIS
When a database slows down, JAS-MIN helps identify what changed in its activity at the same time. That gives us a list of places to investigate: specific SQL statements, disk waits, locks or CPU usage.
The order of that list is not enough. We also examine the scale of each issue and when it occurs: all the time, under heavy load or during a single spike. Then we decide what deserves attention first.
A lead in the data. Evidence in the database.A relationship between measurements helps locate a suspect. We check the diagnosis against actual queries, execution plans and sessions before proposing a change.
The result for your application matters.The goal is a shorter response time or higher throughput. Recommendations include a way to measure whether the change delivers the expected result.
02 / FOLLOW THE EVIDENCE
We start by asking where the time goes — and why. This is the path from a bundle of reports to a decision you can justify.
What slows down, when, and who is affected? We agree the audit goal, representative periods and data scope. We check snapshot coverage: a missing measurement is not evidence of a healthy system.
JAS-MIN organizes AWR or STATSPACK reports and exposes trends, anomalies and relationships. We examine normal operation, busy periods and rare spikes. An average can do a great job of hiding a terrible Monday.
We connect signals to SQL, execution plans, wait events, statistics and available logs. When needed, we agree additional measurements. Evidence separates a hypothesis from a cause — not a confident tone of voice.
You receive findings, evidence and priorities, plus risks and verification criteria for recommended changes. Implementation and environment testing are agreed separately. A report is not permission to modify production.
03 / WHAT’S UNDER THE HOOD
TIME / WORKLOAD
DB Time and DB CPU, workload, waits, I/O, instance and segment statistics. We separate periods and instances: one combined result can hide an important difference.
SQL / ENGINE
SQL cost and frequency, execution plans, child cursors, latches and mutexes — where the data supports assessment. If a plan or measurement is missing, we say what else is needed.
PATTERNS / OUTLIERS
Anomalies, correlations and regression models help prioritize leads. We compare typical workload behavior with the tail of the distribution. A model result is not a promise of seconds or money saved.
04 / DELIVERABLES, NOT DECORATIONS
For the decision-maker: the problem, impact and order of actions. For the technical team: detail, linked evidence and clear limits to what is known. No need to guess what the author meant.
05 / AI ASSISTED. HUMAN ACCOUNTABLE.
Local calculations first, then a compact evidence package with its limitations. Optionally, AI can request specific evidence through MCP instead of drowning in thousands of reports. JAS-MIN helps ask better questions. We remain responsible for interpreting the answers.
We agree the data scope, transfer method and any use of an external model with you. JAS-MIN can analyze local reports; AI integration is optional. We review collected material for sensitive information — a masking setting does not replace that review.
NEXT COMMAND / LET’S TALK